Vache prompts. Claude codes.How it works
The engine

How it works.
One loop, end to end.

Most of this site is output. This page is the engine that produces it — a single closed loop where questions become autonomous research, findings link into a knowledge graph, synthesis finds the cross-domain connections, and the patterns that prove out get written back as rules. Each pass makes the next one sharper.

The engine · how it works

One closed loop. It improves itself every cycle.

Questions go in, autonomous agents research them, findings link into a knowledge graph, synthesis finds the cross-domain connections, and the patterns that prove out become rules that sharpen the next cycle. Hover or tap a stage to see what it actually does — annotated with live counts from the vault, not hand-typed.

Stage 1 of 6

Intent

Everything starts as raw input — real prompts, project goals, and open questions. These are the seeds the engine pulls through the rest of the loop.

8,270 open questions

The center is the nervous system — a neural router and a heartbeat scheduler keeping 170 tasks and 2 autonomous loops in step. The same engine is applied across 12 industries and 19 shipped apps.

See it applied Counts are generated from the live vault

Data as of 2026-06-03

Drive the router

One step of the loop, live in your browser. Fifteen specialized agent profiles (“neurons”) compete to handle each task; the winner loads its own context and conventions. Type anything and watch them activate.

Mode

The vault auto-detects mode from your wording. Flip it manually to watch the neuromodulators re-weight whole classes of neurons.

Routes to Debugger

Debugger3.134
typeerrorcrashingwhy does
Infra Ops0.544
heartbeat
Perf Optimizer0.056
caching~
Linesheet Coder0.000
Project Coder0.000
Reviewer0.000

What you're driving is the deterministic keyword tier of the real router, running entirely in your browser — the exact scoring from neural-router.py (task-type keywords count 3×, domain nouns 1× with diminishing returns, negative-keyword penalties, per-mode neuromodulator gains, and a “task-type beats domain noun” pass). The vault reports this tier at ~87% on its own. The live system (~98% reported) adds a Claude Haiku tie-break for ambiguous cases, plus cross-message hysteresis and a learning loop — those run server-side and aren't in this replica. More on the engine.

What it is — and isn't

A workflow, not a model

Nothing here is a new model or a training run. It's off-the-shelf models — Claude, GPT, and small local ones — wired into a loop with persistent memory, so work compounds across sessions instead of evaporating.

Markdown is the substrate

Every note, rule, and bit of state is plain text in an Obsidian vault. No proprietary store, no lock-in — any tool (or agent) can read and write it. That's what makes the graph and the self-improving rules possible.

Honest about the seams

The loop is real and running, but it isn't magic: agents make mistakes, synthesis produces duds, and rules need pruning. The system's edge is that each failure becomes a rule, so the same mistake gets harder to repeat.

Want to see what the loop has shipped?

© 2026 Vache Sarkissian·Built with Claude Code